Anthropic Expands Claude Access for Vetted Cyber Teams as Glasswing Finds 129,000 Flaws
Anthropic is expanding its Cyber Verification Program, allowing vetted cybersecurity professionals to test AI models with fewer restrictions. The initiative has identified over 129,000 software vulnerabilities.
Intelligence analysis by Gemini 2.5 Flash Lite
Anthropic's expanded Cyber Verification Program (CVP) grants cybersecurity teams access to its AI models, including Claude Opus 5.5, with reduced safeguards. This program, building on Project Glasswing, has uncovered a significant number of software vulnerabilities, aiming to equip defenders with AI capabilities for security.
Imagine AI is like a super-smart detective that can find hidden problems in computer programs. Anthropic is letting trusted security experts use their best detective AI, Claude, with fewer rules. This detective AI has already found over 129,000 hidden computer bugs, helping people fix them before bad guys can use them.
Analysis
Project Glasswing
Anthropic's Project Glasswing has been instrumental in identifying a substantial volume of software vulnerabilities. Between April and July 2026, the initiative verified at least 129,000 flaws, with an additional 5,500 found through open-source scanning between April and October 2026. A significant portion of these, over 33,000, have been classified as critical or high-severity. Anthropic suggests that these figures are likely an undercount, potentially being at least five times higher based on survey data from a subset of partners. This highlights the sheer scale of vulnerabilities that can be uncovered with advanced AI assistance, underscoring the ongoing need for robust security practices.
Cyber Verification Program (CVP)
The expansion of Anthropic's program, now formalized as the Cyber Verification Program (CVP), introduces three distinct access tiers: Defense Access, Red Team Access, and Specialized Access. These tiers cater to different cybersecurity functions, from incident response and vulnerability analysis to authorized penetration testing. The Specialized Access tier, with the fewest safeguards, is reserved for a select group of organizations. Evaluations using CyScenarioBench demonstrate a significant reduction in AI model safeguards blocking tasks, particularly in the Red Team Access tier, which performed identically to models with no safeguards applied. This controlled release of AI capabilities aims to empower defenders by providing them with tools that mirror those potentially used by malicious actors.
Vulnerability Exploitation and AI Patching
While AI models like those from Anthropic are proving effective at discovering vulnerabilities, the actual exploitation rate in the wild remains relatively low. Research indicates that only a small percentage of discovered flaws have been actively exploited. This suggests that AI-driven vulnerability discovery, while lowering the barrier, does not automatically translate to immediate widespread exploitation. Furthermore, the article touches upon the emerging challenge of AI-generated code patches themselves introducing new security risks, as seen in tests by Veracode where a significant portion of AI code generation tasks introduced vulnerabilities. This underscores the need for careful validation of AI-assisted security solutions.
Key points
- Anthropic is expanding access to its AI models for vetted cybersecurity professionals through the Cyber Verification Program (CVP).
- The initiative, including Project Glasswing, has identified over 129,000 verified software vulnerabilities, with many classified as critical or high-severity.
- The CVP offers tiered access, including options with significantly reduced safeguards for specialized testing.
- While AI aids vulnerability discovery, the rate of exploitation in the wild for these flaws remains low.
- AI-generated code patches can introduce new security risks, highlighting the need for careful validation.
The expansion of AI access for cybersecurity teams could significantly accelerate the identification and remediation of software vulnerabilities, leading to more secure digital systems. By equipping defenders with advanced AI tools, Anthropic aims to proactively address threats and bolster overall cybersecurity resilience.
There's a risk that the AI models, even with reduced safeguards, could be misused or that the vulnerabilities they discover are not effectively patched, potentially creating new attack vectors. Additionally, AI-generated code patches might introduce their own security flaws, complicating remediation efforts.
